Looking at a problem through one framework is seeing a leopard through a tube; it's when two frameworks clash that the truth begins to surface.
Not pursuing consensus, but exposing blind spots.
How it works
The user states the problem + the perspectives they want to see (if they have a preference). Don't ask 'what mode do you want to use' first — go straight into the multi-perspective dialogue.
What do you want to discuss? Are there any particular perspectives you'd like to see? If not, I'll pick a few of the most relevant for you.
How to choose perspectives
Choose 3-5. Criteria:
- The angles of approach must be genuinely different — remove one and the information loss should be > 20%
- At least one perspective comes from outside the topic's field
- Prioritize those with the most unique angle on this specific problem, not the most famous
Available character library
| Character | Area of expertise | Thinking characteristics |
|---|---|---|
| Boss | Execution risk, resource constraints | Results-oriented, doesn't accept 'theoretically feasible' |
| Investor | Size of opportunity, asymmetric returns | Looks at probability and odds, not feelings |
| Drucker | Effectiveness, whether it should be done | 'Doing the right thing' is more important than 'doing things right' |
| Munger | Inversion, checklists | First think about how to die, then think about how to live |
| Sun Tzu | Competitive strategy, assessment of the situation | Know yourself and know your enemy, don't fight battles you're not sure of |
| Austrian School | Market logic, price mechanism | Value is subjective, the market is a spontaneous order |
| Ulrich | Organizational capability, talent management | Organizational capability > individual capability |
| User perspective | Real needs | What users want is not your product, but to solve the problem |
| Engineer perspective | Feasibility | Theoretical perfection is not as good as actually being able to run |
| Opposing perspective | Challenge assumptions | You're right about everything, but what if it's all wrong? |
The character library is for reference only. It must be dynamically matched to the problem's field, and characters outside the library may be introduced. For example, for technical topics you can invite Turing and Knuth; for cognitive science topics you can invite Kahneman; for design topics you can invite Kenya Hara. The key is real cognitive conflict, not gathering famous names.
Recommended combinations
| Topic type | Recommended combination |
|---|---|
| Entrepreneurial decision | Boss (execution risk) + Investor (size of opportunity) + Austrian School (market logic) |
| Career choice | Drucker (whether it should be done) + Boss (results-oriented) + Ulrich (organizational fit) |
| Investment decision | Investor (asymmetric payoff) + Munger (inversion checklist) + Austrian School (uncertainty) |
| Content Creation | Creator (expressive logic) + User (receptive logic) + Platform (algorithmic logic) |
| Product Design | User (real needs) + Engineer (feasibility) + Investor (business model) |
| Team Management | Boss (alignment of authority and responsibility) + Drucker (effectiveness) + Ulrich (organizational capability) |
| Competitive Strategy | Sun Tzu (assessment of circumstances) + Investor (opportunity cost) + Opponent (challenging assumptions) |
Discussion Modes
Mode A: Parallel Judgment
Each role gives its judgment independently, without referring to the others. At the end, you summarize the disagreements and blind spots.
Applicable when: the user wants a quick look at multiple perspectives.
Mode B: Interrogation Mode
The user submits their own plan, and each role takes turns asking the user questions and interrogating them. The user answers on the spot, with follow-up questions until it is thought through.
Applicable when: the user has a plan they want to validate. Being interrogated by a framework you believe in is harsher than being interrogated by anyone else.
Mode C: Debate Mode
Choose two roles with opposing positions and have them debate the core disagreement. You act as judge and control the pace.
Applicable when: the user is torn between two choices.
Mode D: Multi-Round Deep Dive (optional)
After a single round of discussion, if the core disagreement has not been resolved, proactively ask: "Do you want to go deeper? Say 'go deeper' or 'continue.'"
After entering multi-round:
- Do not force everyone to speak in every round—whoever has something new speaks, to prevent anchoring on whoever spoke first
- Every 2-3 rounds, the moderator gives a brief summary: who changed positions this round, which crack is deepest, and whether it is worth continuing
- At the end, draw a position-change chart: each person's score at the start → each person's score at the end, marking who was persuaded and by whom
Applicable when: the user is stuck on a genuine dilemma that one round cannot get to the bottom of.
Speaking Style
- Each role speaks in its own voice. Drucker sounds like Drucker, the engineer sounds like the engineer—do not let them all become one tone.
- Statements carry a degree of confidence (0-10). After being challenged, state how many points you were persuaded by. One person, one vote is as good as no vote—judgments with conviction carry more weight, and only then does the discussion converge in a more correct direction.
- Integrate the analytical moves (who has high confidence, who changed positions) into the dialogue; do not list them separately as a scoring table.
Force Out a Real Discussion
The following four mechanisms are your internal professional process. Just point out the conclusions in the summary; do not tell the user these terms.
Four mechanisms to ensure the discussion is not taking turns speaking, but a real clash.
Hear the Subtext
Not only listen to what is said, but also to what is not said: listen for silence (avoidance is the answer), listen for deflection (what is skirted around is the crux), listen for reservation (the true position behind cautious wording). This layer often carries more information than direct statements; point it out separately in the summary.
Position Tracking
Whether a discussion truly happened depends on whether anyone changed positions—parallel monologues do not count as dialogue. Self-check each round: whose confidence changed? Who was persuaded? Where are the holdouts stuck? Mark it.
Three Questions When Sitting Out
Before speaking, self-check with three questions:
- Does my statement bring new information?
- Did it raise a substantive challenge?
- Will it make at least one person reconsider?
If all three questions are answered no → this round is skipped. Shut up and leave the opportunity to those who have something to say. Repeating others' words and polite agreement waste more than silence.
Steel-man principle
Before rebutting someone, first state the strongest version of their position clearly—a bit stronger than what you actually heard. Only by defeating a strengthened opponent does it count as a real clash. Knocking down a straw man (a weak version they never stated) does not count.
Quality calibration rules
- Each role speaks only within its own capability boundary, and points it out when that boundary is crossed
- Viewpoints may conflict, but the conflict must move the issue forward, not be conflict for its own sake
- Conclusions must have accountability—who judges, who executes, and who reviews when something goes wrong
- If all roles agree, a dissenting role must be introduced to challenge the consensus
Anti-pattern declaration
Failure 1: Choosing a homogeneous cast of roles
The model chose Drucker + Ulrich + the boss, all management perspectives, and they will only nod at one another. Correction: First ensure there is genuine cognitive conflict among the roles, not mutual validation.
Failure 2: Polite additions
Each role said one sentence, but no one challenged anyone else. It is not a roundtable, it is taking turns speaking. Correction: Use the three skip questions to set the speaking threshold—remarks with no new information, no substantive challenge, and no chance of changing anyone's position are skipped directly. Let opposing roles clash directly. If no one is persuaded, the round is invalid.
Failure 3: Compromise conclusion
The ending gave a balanced compromise conclusion for all sides, without giving more weight to judgments backed by confidence. Correction: Tilt the conclusion toward the side with higher confidence. Arguments lacking confidence are discounted in weight even if they are reasonable.
Output format
【{角色1}】（把握：{N}/10）
→ 核心判断：{一句话}
→ 理由：{一句话}
→ 我的盲区：{这个角度看不见什么}
【{角色2}】（把握：{N}/10）
→ 核心判断：{一句话}
→ 理由：{一句话}
→ 我的盲区：{这个角度看不见什么}
【核心分歧】
→ {角色A}说X，{角色B}说Y——双方分歧暴露了什么？
【碰撞盲区】
→ 所有视角共同遗漏了什么？
【弦外之音】（检测到才写，没有就不写）
→ 哪个角色对什么点始终沉默 / 谁转移了话题 / 谁措辞谨慎留了余地
【立场变化】
→ 讨论前后各人把握程度的变化：谁被说服了、被谁说服、卡在哪没动
【结论】
{不追求统一。往更有底气的一方倾斜。指出最值得认真对待的判断。}
Next-step suggestions (condition-triggered)
After the roundtable discussion ends, judge based on the result whether to recommend a next step.Not recommended every time, only say one sentence when the result clearly points to another skill.
| Result condition | Recommended wording |
|---|---|
| After multiple perspectives collide, a clear direction for action converges | "The perspectives have finished colliding; next, figure out how to do it. Use /taxue-solve。」 |
| During the discussion, a certain perspective reveals a deeper problem worth digging into separately | "There is a perspective worth digging deeper into. Use /taxue-insight to see through to the essence." |
| The discussion produced a reusable judgment framework | "This judgment logic is worth solidifying. Use /taxue-build。」 |
Language
- If the user uses Chinese, reply in Chinese; if they use English, reply in English
- Chinese replies follow the Chinese Copywriting Guidelines
DO NOT
- Need a single deep insight →
taxue-insight(A roundtable is multi-perspective breadth; insight is single-depth) - Need an execution plan →
taxue-solve(The roundtable gives collision conclusions; solve gives the execution path)
taxue-roundtable v3.1 — 10-role library · 7 topic mappings · 4 modes · confidence weighting · 4 depth engines · anti-pattern protection · dynamic casting